Inspiration
We were inspired by Quiz Bowl, a fun academic competition where quick thinking and broad knowledge are rewarded. We wanted to explore how machine learning could recognize the subject of a question—like science or literature—just from the text, and create a simple, interactive app for anyone to try.
What it does
Our project is an app that acts as a subject classifier for Quiz Bowl-style questions. You simply type in a quiz question, and our model predicts its category (history, science, math, geography, or literature). The app is fast, interactive, and works right in your browser or on your phone.
How we built it
We built a small neural network using PyTorch, with word embeddings and a single linear layer, modeled after classic text classification techniques. The model is trained on trigrams from a set of example quiz bowl questions with labeled categories. For the MVP app interface, we used Streamlit, which lets us combine Python-based AI models and a slick, responsive user interface in a single file, deployable to any laptop or even cloud services.
Challenges we ran into
Training data: Quiz bowl questions cover a huge range of topics, and we learned that a small demo model can mislabel very ambiguous or unusual questions. Model size: We balanced model complexity so it would run instantly in an interactive app. App deployment: Getting a machine-learning model to run live in a browser, with a mobile-friendly interface, was new but made much easier by Streamlit.
Accomplishments that we're proud of
We created a real, interactive demo where anyone can play with AI-powered question classification—on their own device! The app responds instantly and provides readable, user-friendly feedback. We learned to connect backend AI with a modern front end in a minimal codebase.
What we learned
How to implement word embeddings and n-gram neural models for text classification. The importance of cleaning and expanding data in NLP projects. How user experience matters: a usable app is key to communicating our ideas and results.
What's next for Quiz B
Add a much larger and more diverse dataset to train the model so it generalizes to more topics and question styles. Let users submit their own labeled questions and help crowdsource training data. Add more categories (fine-grained subjects). Explore using large language models (LLMs) for either zero-shot classification or richer explanations. Possibly release as a mobile app, or integrate with classroom or competition tools.
Log in or sign up for Devpost to join the conversation.